GDDR6 vs HBM2 What's the consensus?

>>58848069
>GDDR6
High power draw, high latency, potentially as fast as HBM, still slower than HBM2.

>HBM2
Very fast, low power draw and low latency. Smaller memory controller. Requires a complex MCM and is more expensive to implement but allows for much tighter packaging.


Overall, HBM is the future and is just straight up better, but engineering difficulties will keep GDDR in use for a long time to come.

>>58848212
are there gpu processors that can deal with that much bandwidth from hbm2? wouldnt high latency be best if the processor cant handle huge bandwidth, since it will get the smaller amounts of data it can handle faster?

>>58849496
To answer your first question no

GDDR6 because HBM2 is dead end, even AYYMD HOUSEFIRES can't have HBM2 in low end GPUs due to high cost and difficulty with interposers

>>58849523
>HBM2 is dead end
Then why is Nvidia using it in their top-of-the-line HPC compute and workstation cards?

>>58849496
>wouldnt high latency be best if the processor cant handle huge bandwidth
High latency at X teraflops is stil X teraflops wanting to go through the processor.

>>58849523
um... maybe maybe not... hbm2 wasn't ready till late in the year, and if you wanted it before then, you were paying close to full wafer price for it if not more.

once vega hits with hbm2 thats when its ready for mass production, and we will see next gpu cycle

is hbm really that much more expensive?
look at gddr5, they charge us 10$ a gb, is it really THAT much harder to make it on... the fuck is ti called, socket, package, substrate... whatever, is it really that much more expensive to make it part of the die?

I mean honestly, we have no idea what gddr5x gddr6 or gddr5 costs them but we know what it costs us, and how hard is it to mcm?

Look at vega and look at what they are doing server side, it seems amd is going to become an mcm specialist company, so is it really THAT hard to mcm?

We will know next run of gpus, and lets not confuse something, amd is going to be making extremely low end 'run the fucking os' gpus, or at least it seams that way now, along with low end run the fucking game gpu, these are made to keep costs down to a bare minimum, so they may never receive hbm2/derivative memory, but something 150$ or more, that's where we will know.